-
Notifications
You must be signed in to change notification settings - Fork 2.1k
Show PTX and ISA from OpenCL kernel
magnum edited this page Oct 11, 2015
·
5 revisions
- Have JtR save a "binary" (actually PTX) of the kernel. This can be achieved by setting
DUMP_BINARY
environment variable. File name will be kernel's name plus a ".bin" extension, in current directory. ptxas --gpu-name sm_52 foo_kernel.cl.bin -o foo.o
nvdisasm foo.o > foo.isa
Note: CUDA 7.0 ptxas complained about a 352.21 PTX file having "Unsupported .version 4.3; current version is '4.2'" but that can sometimes be worked around simply by editing it 😎
For AMD, use AMD_OCL_BUILD_OPTIONS_APPEND=-save-temps
.